Looking to integrate computing and network domains into a unified artificial intelligence (AI) management system, Arista Networks unveiled a technology demonstration of AI data centers in partnership with Nvidia. This initiative will enable customers to configure, manage, and monitor AI clusters uniformly across essential components such as networks, NICs, and servers, facilitating the creation of optimized generative AI networks with reduced job completion times.

Arista says its collaboration with Nvidia moves it toward establishing a multi-vendor, interoperable ecosystem with coordinated control of AI networking and AI compute.

Recently, I spoke with Pradeep Jothimani, manager of systems engineering at Arista, on the announcement. One of the key points Jothimani made was that there is a tight connection between computing and networks in the age of AI.

“In the front-end network, if you look at compute, it doesn’t have the same strain on the network as the AI workloads,” he told me. “With the AI workloads, the performance of the compute is only as good as the network allows it to be.” With AI, there is a tremendous load on compute, but the network plays a critical role in AI performance.

“As soon as we started working with our customers, we realized that the network can make or break the performance of compute,” he told me. “You can invest in compute. But you don't get the rewards if your network is not designed right.”

This is an essential point for investors and the media to understand. With the rise of AI, Nvidia has become the poster child for the AI era. Unfortunately, most of the press haven’t looked more broadly at AI's impact on other IT segments. I believe Arista will benefit disproportionately from networking for AI compared to the other vendors. I recently dubbed Arista to be a ZK Research “AI Disruptor” and recorded this video with CEO Jayshree Ullal to discuss the opportunity. Arista cut its teeth with high-frequency trading and hyperscalers and has always excelled in areas that require the best possible performance and AI is its next area of focus.

Aiming for uniform controls

Jothimani said that as AI clusters and LLMs expand, the complexity and number of diverse components also tend to increase. However, for a network to be cohesive, companies need to integrate disparate elements like GPUs, NICs, switches, optics, and cables. Once integrated, customers need consistent controls between the AI servers where the NICs and GPUs live and the AI network switches wherever they are. These elements depend on one another for the successful completion of AI jobs, but they operate independently, which can lead to misconfiguration or misalignment.

One example Arista cited is a need for more alignment between NICs and the switch network, which can slow job completion because of the time it takes to diagnose network issues. The company also points to large AI clusters that require coordinated congestion management to prevent packet drops, under-utilization of GPUs, and unified management and monitoring to optimize compute and network resources simultaneously.

Optimizing AI clusters

Arista’s EOS-based agent is at the core of this solution, communicating and coordinating between the network and the host to optimize AI clusters. On Arista switches, EOS can extend its capabilities to directly attached NICs and servers with a remote AI agent, which provides a unified point of control and visibility across an AI data center. The remote AI agent, hosted on an NVIDIA BlueField-3 data processing unit (DPU), also known as a SuperNIC, gives EOS on the network switch the ability to configure, monitor, and troubleshoot network issues. This approach provides consistent end-to-end network configuration and Quality-of-service (QOS), with AI clusters managed and optimized as one.

Some final thoughts

As we were wrapping up, Jothimani made a remark that could seem like a throwaway comment. “It’s networking for AI that can make AI for networking even more valuable,” he said. Think about the telemetry that Arista has. When combined with AI, the implications are very intriguing.

I’ve written about that before. I recently authored this SDxCentral post on “Why networking for AI needs AI for networking.” All networking vendors are now using AI to manage their network infrastructure better, but few have tackled networking for AI, and Arista has addressed both sides of this coin. The partnership with Nvidia should help quell the notion that Nvidia is out to kill all the Ethernet vendors. Some customers prefer Infiniband, and some prefer Ethernet. The key is to give customers a choice and this partnership addresses that.